Audio and speech

Dialogue, TTS, voice, cleanup, separation, translation and localization

• articleWhat do the BBC Casualty accessible-audio trial, a Fraunhofer-linked thesis and Dolby's white paper report about how much viewers adjust dialogue-versus-background balance, and how variable it is?
  Older listeners in one small study (11 listeners aged 66-75, no hearing devices; 2020 bachelor thesis using the Fraunhofer IIS listener pool) chose speech-to-background differences far above the 8.4 LU default: about 15.3 LU with separately produced objects and 13.5 LU with dialogue-separated material, with wide interquartile ranges (11.8-20.8 and 9.8-17.1 LU). The BBC's 2019 Casualty trial let viewers slide between the broadcast mix and a simplified mix; Dolby reports widely differing preferences; SVT chose one fixed Enhanced Speech track. Together they support personal control where separated or object audio exists, with a pre-rendered alternative as the simpler delivery (cross-source synthesis, not a measured comparison).
• articleAt the archive handoff (access file or package to FFV1/Matroska, BagIt and PREMIS/OAIS), what does each primary text guarantee, which package-level information has no home in the archive container, and which records must be written explicitly?
  Integrity checks work at different layers: FFV1 with ec=1 stores a CRC in every slice (MediaConch's sample digitisation policy requires per-slice CRCs), ffmpeg's framemd5/framehash hash decoded frames so two files can be compared by content, and a BagIt 'valid' result only says every manifest checksum matched the stored files. PREMIS 3.0 leaves which events to record to the repository and provides eventOutcome/eventOutcomeDetail for results, so validation outcomes are not automatic archive metadata. Matroska (RFC 9559) can flag hearing-impaired, visually-impaired and text-description tracks, so access roles have a place if a transcode step sets them; the LoC FFV1/Matroska page says timecode support is still in development, so DCP/IMF structure and timecode need separate records (no inspected source maps CPL reels/edit rate/KDM into Matroska).
• articleDoes Chatterbox mark its generated audio, how can I check whether a clip carries that mark, and what does the code actually do?
  Chatterbox's TTS and voice-conversion classes construct Perth's implicit watermarker and apply it before returning audio (the producer reports the same in the multilingual and turbo paths), and the README provides a detection snippet. Robustness claims are the project's own. There is no consent check in the code read. If Perth's implementation fails to import, the class is None, so Chatterbox construction would fail rather than silently skip watermarking (curator code inference, not executed); since the code is MIT, modified copies can omit the call.
• articleWhich audiovisual-media and e-book products and services does Directive (EU) 2019/882 (the EAA) cover from 28 June 2025, and what are the main carve-outs and transition dates?
  The EAA covers consumer terminals used to access AV media services and e-readers (products), and services providing access to AV media services (incl. EPGs) and e-books/dedicated software (services) after 28 June 2025, not AV content itself. The 28 June 2030 transition lets service providers keep using products already in use; it is not a general deadline extension. Microenterprises providing services are exempt.
• articleHow do the engines' docs and source treat sample rate conversion, mix rate and frame-rate divisibility when recording or exporting cinematic audio?
  Godot's docs say the Movie Maker mix rate must divide by the FPS, but the code only warns and sizes each frame's audio block with integer division. At 44.1 kHz and 24 fps that is 1837 instead of 1837.5 samples per frame: 12 samples (about 0.27 ms) lost per second, about 16 ms per minute (curator arithmetic). The 48 kHz default divides evenly at 24/25/30/60 fps. FPS is an integer, so 23.976/29.97 need external retiming.
• articleWhat do Godot, Unity and Unreal audio docs say about keeping an in-engine mix from clipping and about loudness targets?
  Engine docs cover peak safety, not loudness: Godot's hard limiter defaults to a -0.3 dB ceiling and the (soft-clip) limiter to -0.1 dB, so 'add a limiter to Master' does not fix which behaviour you get; no engine page read gives a loudness target or meter, so delivery to a LUFS spec needs external measurement of the exported file.
• articleHow do I get the audio of an in-engine cinematic out as a file in Godot and Unreal, and what format and sample rate will it be?
  Godot Movie Maker records audio non-real-time at a 48 kHz default; the PNG writer writes a stereo WAV whose header declares format code 1 (integer PCM) at 32 bits per sample, while OGV/AVI embed audio. Unreal MRQ WAV parameters are not stated in the pages the producer read.
• articleWhat tense, person, naming and wording rules do the ACB, DCMP and Netflix audio description guides give for writing a description line?
  ACB 2009, DCMP and Netflix v2.5 agree on present tense, active voice, third person (Netflix allows second-person plural for some children's content) and objectivity. On film terms they disagree: ACB advises avoiding 'close-up' or 'fade to black', and Netflix allows common film terms such as 'now in close-up' when needed. On naming, all three describe a character by a visible attribute until the name is established, with exceptions (Netflix allows earlier naming in some cases).
• articleWhat did Ofcom's four live-subtitling quality reports measure and conclude, and what does the NER model count, when the original NER chapter is not available?
  Ofcom's fourth live-subtitling report (2015) measured 98.55% average NER accuracy (97.93% when subtitles above 200 wpm count as errors) and 5.6 s average latency against the 3 s its guidelines recommend; the 98% acceptability threshold came from its external reviewers, not from a stated regulatory requirement.
• articleWhich Apple Music delivery page states a loudness or true-peak limit, and which audio does it apply to?
  For music deliverables, the Asset Guide 5.3.19 gives numeric limits only for immersive audio (Dolby Atmos BWF ADM and 5.1/7.1 LPCM .mov: integrated loudness not above -18 LKFS and true peak not above -1 dB TP per BS.1770-4); stereo and Apple Digital Masters sections give no numeric peak/loudness limit, only an encode-clipping audition rule tied to the Apple Digital Master badge. Applying the immersive numbers to stereo masters is not supported by the guide.
• articleHow does SMPTE ST 2067-2:2020 carry IMSC timed text and fonts in an IMF package?
  SMPTE ST 2067-2:2020 requires an IMF Timed Text Track File to carry one TTML document conforming to the IMSC1.1 Text or Image profile (wrapped per ST 429-5), with non-generic fonts carried as OpenType ancillary resources identified by UUIDv5 derived from the font family name, and defines kinds including subtitle, caption and forced narrative; it sets no characters-per-line or reading-speed limits.
• articleWhat does SMPTE ST 429-2:2023 constrain for timed text in a DCP: fonts, number of subtitle instances, Text and Image elements, tracks?
  SMPTE ST 429-2:2023 requires exactly one LoadFont when Text elements are present, says the font resource should not exceed 10 MB (noting legacy systems may not support more than 640 KB), allows at most two visible subtitle instances with up to six Text or three Image elements each, and limits compositions to one on-screen plus six off-screen text tracks; it sets no characters-per-line or reading-speed limits.
• articleWhat does SMPTE ST 428-7:2014 say normatively about subtitle structure, text limits and font carriage for a DCP subtitle file?
  SMPTE ST 428-7:2014 defines DCDM subtitle XML with percentage positions, an optional LoadFont naming a font only by urn:uuid (mapping to a file out of scope), PNG sub-pictures bounded by the picture, and ruby, vertical and bidi rules; it advises that a Text string fit one line, with multiple lines as sibling Text elements, and sets no characters-per-line, reading-speed or minimum-duration numbers.
• articleWhat does a public studio delivery specification say about handing off Foley (cue sheets, track naming/layout, stems)?
  Netflix's nonfiction M&E guide treats Foley as an optional, often unbudgeted pass cued at the sound or post supervisor's discretion (cloth and props first, footsteps rarely), with married music, effects and separate dialogue stems when there is no Foley; its sound-mix spec v1.6 lists Foley as a labelled object/bed group and expects Foley fill in the 5.1 M&E. No public Netflix text read gives a Foley cue-sheet format or track layout.
• articleWhat does EBU R 37 say numerically about sound and picture timing, and do the EBU-TT-D documents (Tech 3380, Tech 3381) state any reading-speed or line-length numbers?
  EBU R 37-2007 limits end-to-end sound-picture offset at emission to 40 ms sound-early and 60 ms sound-late, with 5 ms early to 15 ms late at each stage, measured with flash-and-blip markers. EBU-TT-D (Tech 3380) sets layout defaults (cellResolution 32x15, linePadding 0c) but neither it nor Tech 3381 gives reading-speed or line-length limits, which must come from style guides.
• articleWhat does AES TD1008 say about true-peak limiting, codec overshoot and sample-rate conversion when preparing masters for lossy streaming?
  AES TD1008 (2021, superseding TD1004) recommends at most -1 dBTP at the input of lossy codecs (high-rate coders around 256 kbps may tolerate -0.5 dBTP), a true-peak limiter as the last element before the codec, and notes downward SRC can create overshoots a limiter placed before it cannot catch; its distribution table targets speech at -18 LUFS (+1 LU), track-normalised music at -16 LUFS and album-loudest tracks at -14 LUFS (+0.2 LU).
• articleWhat does Apple Digital Masters (formerly Mastered for iTunes) ask a mastering engineer to deliver and check for AAC streaming: bit depth, sample rate, headroom, SRC and clipping?
  Apple's Digital Masters brief asks for 24-bit sources at the highest native rate (no upsampling or bit-padding), describes mastering-quality SRC to a floating-point intermediate before AAC encoding, recommends at least 1 dB headroom because playback oversampling and encoding can clip material that shows no PCM overs (checked with afclip), and says one master feeds both AAC and ALAC. It gives no dBTP or LUFS target.
• articleWhat subtitling, audio description and signing percentages does Ofcom's Code on Television Access Services set for UK-regulated TV channels, and how are channels and affordability decided?
  Ofcom's 2024 access code sets statutory ten-year targets of 80% subtitling (90% Channel 3 and 4; 100% BBC), 10% audio description and 5% signing as percentages of non-advertising programme hours per delivery platform, with interim targets by anniversary (e.g. year 1 10/2/1%, year 5 60/10/3%); Ofcom's July 2025 notice lists 83 channels required in 2026 (82 at Level One). The Code does not cover on-demand services (producer reading of the 2024 statement).
• articleWhat do the FCC closed-captioning quality rules in 47 CFR 79.1 actually require, and do they set numeric timing or line-length limits?
  FCC caption quality standards in 47 CFR 79.1(j)(2) require captions to be accurate, synchronous, complete and appropriately placed (readable speed; not blocking important visual content), with a de minimis allowance and case-by-case factors for live programming; the section read contains no numeric speed, line-length or timing limits.
• articleWhat do the W3C MAUR and UAAG 2.0 ask of media players for caption customisation, keyboard operation and sign-language windows?
  W3C's MAUR (2015 Note) lists player requirements: configurable caption fonts, sizes, colours and background opacity, caption track toggling, keyboard operation of all functionality, and sign-language video as a track or file shown as picture-in-picture, overlay or parallel video with a user toggle. UAAG 2.0 (2015 Note) adds that alternatives must not obscure controls or video, captions are user-configurable, and alternatives can be resized to at least 50% of the viewport and repositioned. Both are Notes, not Recommendations.
• articleWhat audio and asset deliverables does Netflix require from a dubbing studio besides the script, and what public detail exists on dub cards?
  Netflix's premix spec asks for a mono offline-bounced WAV of all premix dialogue with group walla attenuated by 10 dB and a -6 dB true-peak limiter on dialogue or the bounce bus, and the dialogue editorial session as a store-only ZIP named with 48k24b; dub-card requirements could not be read (page is a JavaScript shell; the article API returns not found).
• articleWhat script format and content rules does Netflix publish for the dubbing script (as-recorded script) a dubbing studio must deliver?
  Netflix's as-recorded dubbing script is delivered in TTAL (direct upload or its Script Authoring tool) and validated; events are timed to the first and last spoken word, split at pauses of about two seconds, with WALLA for indistinct crowd dialogue; delivery is due three calendar days before the secondary audio (changed from five on 2024-09-17).
• articleWhich WER variants and text normalisations change ASR subtitle scores, and is COMET usable for subtitle translation evaluation?
  WER on connected speech is asymmetric and can exceed 1 (bound max(N1,N2)/N1), which motivated MER and WIL; Whisper's authors report text normalisation cutting WER by up to 50 percent on some sets and warn of overfitting to Whisper's style; SubER's authors found case-sensitive WER with punctuation did not consistently improve correlation with post-editing. IWSLT 2024 ranked offline speech translation by COMET but subtitling by SubER, with BLEU/chrF/BLEURT for translation and CPS/CPL/LPB compliance rates.
• articleWhat rendering failures for RTL, ruby, vertical and text-combine timed text are documented in public renderer issue trackers and test suites?
  The imscJS tracker records concrete RTL/CJK failures: per-character spans broke Arabic joining in WebKit (109, closed), sheared vertical Japanese text was cropped by its region (260, closed), rubyReserve broke on Chrome after a switch to CSS ruby-text-container (289, closed 2026-10-10), and ruby style-resolution pruning (162) and xml:space/rubyAlign (185) remain open. One renderer's tracker; a sample, not prevalence.
• articleWhat do IMSC/TTML2, WebVTT and the Netflix Japanese guide say about ruby annotations, vertical text and tate-chu-yoko in subtitles?
  IMSC 1.3 permits ruby (via ruby, rubyAlign-minimal, rubyPosition, rubyReserve), text combination and vertical writing; WebVTT has ruby spans, vertical growing-left/right writing and CSS ::cue text-combine-upright but no cue setting for it. Netflix's Japanese profile sets 13 (horizontal) and 11 (vertical) full-width characters per line, half-width counting 0.5, up to 4 characters per second, and ruby/bouten placement by orientation and line.
• articleHow do I score machine-generated subtitle files (not just text) against a reference, what does SubER measure, and how far has it been validated?
  SubER scores a subtitle file against a reference including text, line breaks and timing (SRT input; Apache-2.0 code); its authors give a rule of thumb (<20% very good, >40-50% bad). In IWSLT 2024 human judgments for translation-focused DA agreed more with BLEURT than SubER (producer reading), so its validation depends on the evaluation target.
• articleHow are forced narrative subtitles represented and chosen at playback in IMSC, WebVTT and on Netflix, and what language-specific conventions apply?
  IMSC 1.3 marks forced content with itts:forcedDisplay and a player-side displayForcedOnlyMode; Netflix shows FN only when subtitles/CC are off, so full subtitle and SDH files must themselves contain the FN content. WebVTT (CRD read by producer) has no forced-narrative concept; language guides add marking rules (quotes, hyphens, italics).
• articleWhat does an independent evaluation measure when it compares MFA, WhisperX and MMS word alignment against manual alignments, and what are its stated limits?
  Against manual alignments on TIMIT and Buckeye (words correctly recognised only), MFA places more word boundaries within 10-100 ms than WhisperX or MMS; on Buckeye WhisperX reaches only 77.4% within 100 ms. The authors note their Buckeye result differs from the original MFA paper (possible long-utterance drift).
• articleHow does WhisperX produce word timestamps, under which separate licences (code, alignment models, diarization), what accuracy does its paper report and on what, and what failure cases does the repository document?
  WhisperX (BSD-2 code) transcribes with Whisper and aligns words with a language-specific wav2vec2 model (English default torchaudio WAV2VEC2_ASR_BASE_960H); words with characters outside the model dictionary (numbers, currency) get no timing, overlapping speech is weak, and diarization needs a gated pyannote model with a user agreement.
• articleUnder what separate terms are the Montreal Forced Aligner code and its English acoustic model released, what training data and performance caveats does the model page state, and what does the model page say about its own accuracy metric?
  MFA code is MIT while the English acoustic model v3.0.0 is CC BY 4.0 and lists its training corpora; its 0% WER/CER is on training data (not held-out); the page warns about demographic bias and recording consent. Pronunciation dictionaries carry separate licences (not read).
• articleWhat do public summaries of the 2023 SAG-AFTRA TV/Theatrical AI provisions say about digital replicas of voice and likeness, synthetic performers and consent, and what primary text is unavailable?
  A law-firm summary (2023-11-13, before ratification) describes the 2023 SAG-AFTRA TV/Theatrical AI terms: consent requirements for employment-based and independently created digital replicas and digital alteration (arbitration, monetary remedies), and notice/bargaining for synthetic performers. Union primary texts were not readable.
• articleWhat does the Tennessee ELVIS Act text say about voice, including simulated voice and voice-cloning tools, and how does the adopted amendment differ from the bill as introduced?
  The ELVIS Act amendment defines protected voice to include a simulation, creates liability for knowingly making an unauthorised voice/likeness available and for distributing tools whose primary purpose is producing a particular, identifiable individual's voice or likeness (narrowed from the introduced text's 'an individual's ... without authorization'), lists First Amendment fair uses, and lets exclusive recording-artist contract holders sue.
• articleWhat does Article 50 of the EU AI Act say about marking synthetic audio and disclosing audio deep fakes, who is addressed, and what exceptions does the text state?
  As reproduced on a secondary site, AI Act Article 50 requires providers of systems generating synthetic audio/image/video/text to mark outputs machine-readably (except assistive standard editing), requires deployers to disclose deep fakes, and limits that disclosure for evidently artistic or fictional works to a form that does not hamper enjoyment.
• articleHow are timed text tracks carried and validated in IMF packages according to public sources (ST 2067-2 versions, Netflix profile, Photon)?
  Netflix Photon validates IMF timed-text track files by mapping kind (0x13), UTF-8 UCSEncoding and an IMSC1 profile NamespaceURI; Netflix's own IMSC 1.1 text profile is currently valid only for Japanese timed text. ST 2067-2 text itself was not read (ZIP-only).
• articleHow do Interop and SMPTE ST 428-7 subtitle files differ inside a DCP, and what do public tool checks and delivery rules encode for each?
  As encoded in two open validators, Interop and SMPTE DCP subtitles differ in file form (.xml vs .mxf), root element and timing units, and SMPTE text subtitles require exactly one LoadFont while Interop does not; libdcp checks main-subtitle line lengths at 52 (warning) / 79 (error) characters and closed captions at 32. The validators classify some rules differently (libdcp warning vs ClairMeta error for the 4 s first-subtitle rule).
• articleHow do DCMP, Netflix English (USA) and BBC guides handle music descriptions and song lyrics in captions?
  Lyrics marking differs: Netflix ♪ at start and end of each subtitle plus italics; BBC # at the start of lyric subtitles (centred), mood music labelled only when required; DCMP marks nonessential background music with a ♪ icon top right and never at the expense of dialogue. Only Netflix requires italic lyrics.
• articleHow do DCMP, Netflix English (USA) and BBC guidance write sound-effect descriptions, and where do they differ?
  All three caption only sound that matters and describe sounds rather than actions; DCMP is most prescriptive on form (lowercase bracketed source+sound, optional onomatopoeia line, participle vs third-person tense), BBC uses white caps on a separate line, Netflix lowercase brackets with detailed descriptions for plot-pertinent sounds.
• articleHow should speaker identification be written in SDH/captions, and where do DCMP, Netflix English (USA) and BBC guidance disagree?
  All three guides identify speakers only when the viewer cannot; devices differ: DCMP puts the parenthesised name on its own line, Netflix wants the ID on the same line as the dialogue, and BBC prefers colour (narrator yellow) with ttm:agent metadata in EBU-TT.
• articleWhat validity rules and error cases does Unicode MessageFormat 2 (UTS #35 Part 9) define for .match plural/ordinal/gender selection and unresolved variables?
  Unicode MessageFormat 2 (CLDR text at 7cdeb9e7): a .match message is valid only with matching key/selector counts, a catch-all variant, annotated selectors and unique variants; violations are named data-model errors; unresolved variables fall back to '$name'; the select option must be a literal so messages stay translatable.
• articleHow do I write a translatable dialogue line in Project Fluent that varies by count, gender or grammatical case?
  In Project Fluent a translatable line varies by count, gender or case through select expressions with a required default variant (*): numbers match exact keys or plural categories, formatting options can change the category, parameterised terms select grammatical case, and term attributes like .gender drive agreement.
• articleWhat does RFC 9639 FLAC allow for channel count, sample rate, length and metadata, and how does it carry a non-default channel layout or WAV chunks?
  RFC 9639 FLAC: 1-8 channels, 1-1048575 Hz, 4-32 bits; 36-bit total-sample field (0 = unknown) instead of a RIFF-style byte limit; metadata blocks capped at 2^24 bytes; non-default channel layouts need a Vorbis-comment channel mask and are then not streamable; foreign RIFF/W64/AIFF chunks only via registered APPLICATION blocks (no bext equivalent).
• articleMy WAV is over 4 GB or multichannel with ADM metadata: should I write RF64 or BW64, and how does the file say which it is?
  BW64 (BS.2088-2) and RF64 (EBU 3306) share the ds64/64-bit-size mechanism and JUNK reservation; the EBU itself directs implementers to BS.2088 exclusively; BW64 removes RF64's 18-position channel mask and carries ADM via chna/axml; the top-level ID, not the .wav extension, distinguishes them.
• articleWhat file format, level, peak and noise-floor numbers does ACX state for audiobook submissions, and how do I check an audiobook chapter file against them?
  ACX (2023 snapshot) states RMS-based spoken-word limits (-23 to -18 dB RMS, -3 dB peaks, -60 dB RMS noise floor), 192 kbps+ CBR 44.1 kHz MP3, uniform channel format (mono recommended, no joint stereo), and human narration only; it does not use LUFS or true peak.
• articleWhat licence do the Valentini/Voicebank-DEMAND item and the OpenSLR pages for MUSAN, LibriTTS, LibriTTS-R and Hi-Fi TTS state, and what do they say about training?
  OpenSLR pages for MUSAN, LibriTTS, Hi-Fi TTS and LibriTTS-R each state CC BY 4.0 at page level (byte-identical to the producer pins).
• articleWhat are the practical SCC/CEA-608 limits (columns, rows, timing, capacity) that public tools document?
  pycaption (ad2ddba8) writes SCC for CEA-608 pop-on, roll-up and paint-on modes on a 32-column, 15-row grid, wraps lines at 32 characters, caps a caption at 80 tokens including 8 control codewords, supports drop-frame and non-drop-frame timecode, and times one codeword at 1/29.97 s (static reading).
• articleWhat do Xbox Accessibility Guideline 104 and the Game Accessibility Guidelines require for subtitles and captions in games beyond linear-video rules?
  Xbox Accessibility Guideline 104 asks for subtitles for all spoken content with speaker identification and spatial cues, independent on/off toggles for caption types such as dialogue, background dialogue and ambient sounds, lines of no more than about 40 characters, at most two lines, and text scalable to at least 200% of the default; the Game Accessibility Guidelines give 40 characters per line, two lines (three exceptionally) and a 46px-at-1080p minimum size.
• articleHow did Netflix's published M&E true-peak wording change across archived versions of the Sound Mix Specifications, and where does a 2.0 stereo M&E rule live now?
  Across archived Netflix Sound Mix Specifications (v1.1 2019, v1.4 2022, v1.5 2023, v1.6 2026), every stated M&E ceiling is -2 dB true peak (+18 dBu / -2 dBFS over a -20 dBFS reference); the 5.1 M&E section stated it in 2019 but not in v1.4 onward (by reference via the Atmos rerender rule), the 2.0 stereo M&E section persists in v1.4 although its change log says it was removed in 2020, v1.5 records its removal in 2023, and v1.6 removed Atmos true-peak language.
• articleWhat true-peak and loudness limits do Netflix's public sound specifications state for a Music and Effects (M&E) mix?
  In the Netflix Sound Mix Specifications v1.6 (archived 9 Aug 2026), the 5.1 M&E section states no numeric peak or loudness value of its own; M&E peak compliance is by reference to the Netflix Peak specification, -2.3 dBFS limiter settings are recommendations, and the theatrical 0 dB true-peak allowance belongs to a different deliverable.
• articleDoes natural speech really have the picture (mouth movement) ahead of the sound by about 150 ms, and what offset range should a generated talking-head and its separately generated voice be expected to show?
  Schwartz and Savariaux (PLoS Comput Biol 2014) dispute the "visual speech leads by 100-300 ms" claim: for chained syllables they measured audio-visual asynchrony between about 20 ms audio lead and 70 ms audio lag, and extrapolate 40 ms audio lead to 200 ms audio lag for richer speech. Large picture leads occur at utterance starts or isolated syllables (preparatory gestures). Their data come from a simple French /a/-context corpus filmed at 50 Hz, so timing precision is about 10-20 ms. Love et al. and others still cite the 100-300 ms figure as background, so the sources conflict on the natural lead.
• articleWhat do the open sources say about the gwangdae narrator-singer's craft criteria and the tension between improvised adaptation and a fixed text, and which failure modes (relevance, freezing, over-entertaining) follow for a solo screen or audio storyteller?
  A 19th-century theorist (Sin Jaehyo, via the thesis) named four conditions for a gwangdae: appearance/presence (inmulchire), narration (saseolchire: dramatic, literary, clear to listeners), vocal quality (deoneum) and gesture with props (neoreumsae). The same thesis shows a modern competition where judges weighed original essence, creativity and entertainment, and an expert warned that a filial-sacrifice story may not move modern audiences. Yates-Lu shows the opposite pressure: preservation systems fix lineages, and scholars dispute whether a stabilised version is the 'original', with one scholar arguing differences simply reflect a narrative not yet stable. Together they frame an open tension between adapting to the audience and preserving a text; none of the sources tests either for audio or screen use.

More